Policy Transfer in Reinforcement Learning with Domain Adaptation using Transition Probability

Abstract

Reinforcement learning is drawing increasing attentions in real world applications.Since it often takes enormous cost to learn the agent in the real world environment (called target task), pre-training in a low-cost environment such as a simulator (called source task) is gathering attention. In this paper, we focus on the situation where the source and target tasks are different only in the form of state observation. Our proposed method trains encoders mapping state observation to latent representations, and trains a policy that receives a latent representation and output an action.We utilize the transition probability to learn latent representations robust to changes in the form of state observation.This enables transferring the policy learned in the source task to improve the performance in the target task.Experiments show that our method can achieve higher performance when the number of interactions in the target task is limited.